CSCI 6461 Computer Architecture Branch Prediction Instructor M

CSCI 6461: Computer Architecture Branch Prediction Instructor: M. Lancaster Corresponding to Hennessey and Patterson Fifth Edition Section 3. 3 and Part of Section 3. 9

Reducing Branch Costs • The frequency of branches and jumps demands that we also attack stalls arising from control dependencies • As we are able to add parallel and multiple parallel units, branching becomes a constraining factor • On an n-issue processor, branches will arrive n times faster September 2012 2

Review of a Branching Optimization Branch destination and test known at end of third cycle of execution September 2012 Branch destination and test known at end of second cycle of execution Instruction Level Parallelism 3

Dynamic Branch Prediction • Branch prediction buffer – Simplest scheme – A small memory indexed by the lower portion of the address of the branch instruction • Includes a bit that says whether the branch was taken recently or not • No other tags • Useful only to reduce the branch delay when it its longer than the time to compute the possible target PCs • Since we only use low order bits, some other branch instruction could have set the tag – The prediction is a hint that is assumed to be correct, if it turns out wrong, the prediction bit is inverted and stored back September 2012 4

Dynamic Branch Prediction • Branch prediction buffer is a cache • The 1 bit scheme has a shortcoming – Even if a branch is almost always taken, we will usually predict incorrectly twice, rather than once, when it is not taken • Consider a loop branch that is taken nine times in a row then not taken. What is the prediction accuracy for this branch, assuming the prediction bit for this branch remains in the prediction buffer – Mispredict on the first and last predictions, as the loop branch was not taken on the first one as is set to 0. Then on the last loop it will not be taken and the prediction will be wrong again. – Down to 80% accuracy here September 2012 5

Dynamic Branch Prediction • To remedy this situation, 2 bit branch prediction schemes are often used. A prediction must miss twice before it is changed. • A specialization of a more general scheme that has a n-bit saturating counter for each entry in the prediction buffer. With n bits, we can take on the values 0 to 2 n-1. When the counter is >= ½ of its max value, branch is predicted as taken • Count is incremented on a taken branch and decremented on a not taken one • 2 bits work almost as well as larger numbers September 2012 6

The States in a 2 Bit Prediction Scheme September 2012 7

Branch Prediction Buffer • Implemented via a small special cache accessed with the instruction address during the IF pipe stage, or as a pair of bits attached to each block in the instruction cache and fetched with each instruction. • If the instruction is a branch and if predicted as taken, fetching begins from the target as soon as the PC is known. Otherwise sequential fetching and executing continue. If prediction is wrong the prediction bits are changed as in the state diagram. September 2012 8

Branch Prediction Buffer • Useful for many pipelines • In our five stage pipeline the pipeline finds out whether the branch is taken and what the target of the branch is at roughly the same time as the branch predictor information would have been use (the end of the second stage of the execution of the branch). • Therefore, this scheme does not help for our pipeline • Next figure shows performance of 2 -bit prediction for a given benchmark (between 1 -18% mispredictions) September 2012 9

Prediction accuracy of a 4096 entry 2 -bit prediction buffer September 2012 10

Increasing the size of the buffer does not help much September 2012 11

Correlating Branch Predictors • Branch predictions for integer programs are less accurate • These 2 bit schemes use only recent behavior of a single branch to predict the future behavior of that branch • Look at other branches rather that just the branch we are trying to predict if (aa==2) aa=0; if (bb==2) bb=0; if (aa!=bb){ September 2012 12

Correlating Branch Predictors • MIPS Code L 1: L 2: DSUBUI BNEZ DADD DSUBU BEQZ R 3, R 1, #2 R 3, L 1 R 1, R 0 R 3, R 2, #2 R 3, L 2 R 2, R 0 R 3, R 1, R 2 R 3, L 3 ; branch b 1(aa!=2) ; aa=0 ; branch b 2 (bb!=2) ; bb=0 ; branch b 3(aa==bb) Branch b 3 is correlated with branches b 1 and b 2 – if branches b 1 and b 2 are both not taken then b 3 will be taken since they are equal September 2012 13

Correlating Branch Predictors • Branch predictors that use the behavior of other branches to make a prediction are called correlating predictors or two level predictors. September 2012 14

Correlating Branch Predictors Look at the branches with d = 0, 1, and 2 if (d==0) BNEZ d=1; if (d==1) R 1, L 1 DADDIU R 1, R 0, #1 L 1: ; branch b 1 (d!=0) ; d==0, set d=1 DADDIU R 3, R 1, #-1 BNEZ R 3, L 2 ; branch b 2 (d!=1) L 2; September 2012 15

Correlating Branch Predictors Initial value of d d==0? b 1 Value of d before b 2 d==1? b 2 0 Yes Not taken 1 No Taken 1 Yes Not taken 2 No Taken Possible Execution Sequences • If b 1 is not taken then b 2 will not be taken • A 1 bit predictor initialized does not have the capability to take advantage of this September 2012 16

Correlating Branch Predictors • To develop a branch predictor that uses correlation, let every branch have two prediction bits, one prediction assuming the last branch executed was not taken and another prediction bit that is used the last branch executed was taken. • The last branch executed is usually not the same instruction as the branch being predicted, although this can occur. September 2012 17

1 -Bit Correlation Prediction Bits Prediction if last branch not taken Prediction if last branch taken NT/NT NT/T NT T T/NT T/T T T • This is a 1, 1 predictor since it uses the behavior of the last branch to choose from among a pair of 1 -bit branch predictors • An (m, n) predictor uses the last m branches to choose from 2 m branch predictors, each of which is an n bit predictor for a single branch September 2012 18

(m, n) Predictors • Can yield higher prediction rates than the 2 bit scheme and requires only a small amount of additional hardware We can record the global history of the most recent m branches in an m bit shift register, where each bit records whether the branch was taken or not taken • The branch prediction buffer can be indexed by using a concatenation of the low order bits from the branch address with the m bit global history. That is the address indexes a row in the prediction buffer and the global buffer chooses among them. September 2012 19

Fig 14 September 2012 20

Comparison of Predictors – First is non-correlating for 4096 entries, followed by a non-correlating 2 bit predictor with unlimited entries and finally a 2 bit predictor with 2 bits of global history and 1024 entries September 2012 21

Tournament Predictor for the Alpha 21264 September 2012 22

Fraction of Predictions Coming from the Local Predictor for a Tournament Predictor using SPEC 89 Benchmarks September 2012 23

Branch Target Buffers (Advanced Technique for Instruction Delivery) • Reduce penalty in our 5 stage pipeline – Determine next instruction address to fetch by the end of IF • We must know whether an instruction (not yet decoded) is a branch and, if so what the next PC should be • If at the end of IF we know the instruction is a branch and we know what the next PC should be, we have zero penalty – A branch prediction cache that stores the predicted address for the next instruction after a branch is called a branch target buffer or branch target cache – For the classic 5 stage pipeline, a branch prediction buffer is accessed during the ID cycle. At the end of ID we know the branch target address (computed in ID), the fall through address (computed during IF), and the prediction September 2012 24

Branch Target Buffers • Reduce penalty in our 5 stage pipeline (continued) – Thus by the end of ID we know enough to fetch the next predicted instruction. – For a branch target buffer, we access the buffer during the IF stage using the instruction address of the fetched instruction (a possible branch) to index the buffer – If we get a hit, then we know the predicted instruction address at the end of the IF cycle, which is one cycle earlier than for the branch prediction buffer – This address is predicted and will be sent out before decoding the instruction. It must be known whether the fetched instruction is predicted as a taken branch September 2012 25

Fig 3. 21 A Branch Target Buffer – The PC of the instruction being fetched is matched against a set of instruction addresses stored in the first column; which represent the addresses of known branches. If the PC matches one of these entries, then the instruction being fetched is a taken branch, and the second field, predicted PC, contains the prediction for the next PC after the branch. Fetching immediately begins at that address. September 2012 26

Fig 3. 22 Steps Involve In Handling an Instruction with a Branch Target Buffer September 2012 27